Bar code inclination correction method based on multi-task target detection

Abstract

The invention discloses a bar code inclination correction method based on multi-task target detection, and the method comprises the steps: building a single-stage target detector which comprises a feature extraction basic network and a detection head network which are connected in sequence; inputting the bar code image into the feature extraction basic network, and extracting to obtain a feature map; and inputting the feature map into a detection head network, and carrying out classification and position regression on the feature map to obtain a correction result. According to the method, theinclination degree of the bar code is firstly classified, then angle regression is carried out, the bar code correction difficulty is effectively reduced, the bar code is fused with a single-stage target detector to form a multi-task target detection model, the detection and correction efficiency is improved in parallel, and the decoding accuracy and speed are well improved.

Description

technical field [0001] The invention relates to the technical fields of deep learning and image processing, in particular to a barcode tilt correction method based on multi-task target detection. Background technique [0002] The current mainstream barcode recognition methods are mainly divided into two types: traditional digital image processing and deep learning algorithms. [0003] Traditional image processing methods mainly design features and rules manually, mark the barcode area, calculate the rotation angle, and then obtain the barcode correction result. In general, traditional image algorithms have high requirements for experiments and application environments. In the logistics package sorting scenario, the quality of the barcode image is degraded and the features are not obvious due to uneven lighting, background interference, package distortion, blur, and staining, and the performance of the algorithm is greatly reduced. [0004] In recent years, deep learning te...

AI Technical Summary

Problems solved by technology

Among some existing methods, some are cascaded barcode detection algorithms based on convolutional neural networks. First, the Faster-Rcnn target detection algorithm is used to obtain the barcode area, and then the Maximally Stable Extremal Regions (MSERs) algorithm is used to eliminate the background. noise and detect the direction of the barcode, and finally process the fuzzy barcode area through the adaptive manifold (Adaptive Manifold, AM) filter, the whole algorithm flow is slow due to the multi-stage processing

Others use the YOLO target detection algorithm to locate the barcode area, and then cut and scale the barcode area to a square and send it to an angle correction convolutional neural network to predict the inclination angle of the barcode. However, the features extracted by these two parts of the network have redundancy and are directly regressed. Predicted inclination angle error is too large

Images